You are on page 1of 6

TA-5-1

ADAPTIVE NEURO-CONTROL
FOR SPACECRAFT ATIlTuDE CONTROL
K. KrishnaKumar
Department of Aerospace Engineering
The University of Alabama, Tuscaloosa, AL 35487-0280, USA.
e-mail:kkrishna@ual vm.ua.edu

Abstract transient torque moments. To address some of these


problems, recent studies [6-81 have focussed on other
Spacecraft attitude control is approached as a non-linear alternatives. Reference [6] addresses the treatment of
aaaphve control problem and wuro-control, which combines changing inertia. as an adaptive control problem and uses a
concepts from artificial neural networks and adaptive linear approach for the same. In reference [7l SS control is
control, is investigated as an alternative to linear control approached as a non-linear control problem using
approaches. Two capabilities of neuro-controllers are Lyapunov’s second methood for stability analysis and in
demonstrated using a non-linear model of the Space Station reference [8] a feedback linearhion is used for non-linear
F d o m . These capabilities are: (a) synthesis of robust non- control.
linear controllers using neural networks; ancl (b) aciaptively This paper presents an approach using artificial
maWyingneurocontroller characteristics for varying inertia neural networks (ANN) that will provide both non-hear
. . control and adaptive control capabilities. In the recent past,
S
-C
.- main components of the adaptive neuro-
controllers are an identification network and a controller revived interests in the working of neural networks has
network. Both these networks are trained using the back- brou&t out many new approaches to solving engineering
propagation of error learning paradigm. To ensure problems using mathematical networks which mimic the
robustness of the neuro-controller, an optimally connected workings of neural connections in the brain. Neurocontrol
neural network is synthesized for the identification network. is an application domain of ANN in which ANN concepts
For the on-line adaptive control problem, a new technique are applied to system identification and control. Artificial
using a memory filter for error back-propagation is neural networks have been used in many areas of control
introduced in place of the back-propagation through time problems. In ref- to SS control, references [9]and [lo]
technique. Performance of the non-linear neuro-controllers present mass identification approaches using neural networks
for the three cases listed above are verified using a non- and reference [1 11 presents a neuro-control approach using
linear simulation of the Space Station. Results presented a linearized pitch degree-of-freedom model and shows the
substantiate the feasibiity of using neural networks in robust benefits of using neural networks in adaptive control of SS.
non-linear adaptive control of spacecraft. There are several benefits in using neural networks in
adaptwe non-linearcontrol for space applications. These are:
1. INTRODUCTION
(1) A neuro-controller learns to control a system based on
Spacecxaft attitude control has been examined in the the input-output coupling that exist. Also, neural networks
past using several approaches. In general, these approaches have been shown to extract input-output mapping even out
include classical control techniques [1,2] and modem state of noisecorrupted data. This implies that a decentralized
space techniques [3-51. With the current interest in an controller can be implemented using neuro-controllers with
evolutionaxy approach for constructing a Space Station (SS) direct output feedback. Decentralized control, with direct
in space, there is a need to examine alternate control output feedback, is computationally less demanding and
techniques that can accommodate such evolutionary changes. more suited when using space-quaMed computers that are
Techniques reported in referencesrl-51 use linear approaches limited in their processing power.
for controller designs and make inertia assumptions that are
not valid for most proposed configurations. Also, s m a l l (2) The necessity for a learning-type controller for space
perturbation assumptions made during linearization will be applicationsarises chre to the conditions of uncertainty in the
violated during nominal operations of the SS. environment of operation and due to unmodeled dynamics.
Another substantial deviation from nominal
parameters will occur during space shuttle docking and (3) The possibility of component failure in space structures
general relative mass motion. These contribute si@icantly establishes the requirement for adaptive-type controllers.
to the changes in moments of inertias and also will introduce
The major objective of this study is to highlight the

0-7803-1872-2/94/$4.00 0 1994 IEEE 1153

.- ~ .-
advantages m using neuro-controllers for spacecraft attitude Now, consider the single neuron in the network
control. The Space Station (SS) is chosen as the ideal st"sh0wnmFigure 1. An individualneuronbas many
example for demonstrating these advantages. In what inputs depending on the number of connections. Each
follows, we first present the fundamentals of neuro-control connection to the neumn has a weight associated with it.
relevant to this study and outline a procedure for design and After the net input is calculated, it is converted into an
impkmentation of robust, adaptive neurocontrollers. Next, activation value through a functional relationship. The
we present the non-linear S S model used and show the power of ANN lies ultbin this transfer function. Two types
capabilities of neurocontrol in controlling the pitch attitude of transfer functions are used in this study. These are:
of the SS.These caprabihties are: (a) synthesis of robust non- (a) Sigmoidal function (for all hidden neurons):
liuear concr~llaws using neural networks; and @) adaptively
modifying neuro-controller characteristics using a memory
filter for the back-propagation of error.

2.0 ADAF"IvE NEURO-CONTROL CONCEPTS @) Ramp Function (for all input and output neurons):

Neurocontrol, which combines ANN concepts, f(x) = ax.


optimal control concepts, and adaptive control concepts, is
relatively recent. Naredra [Ill, Werbos [12,13], Barto et In this study, a is chosen to be 1.0. After
al. [14], and many others have made significant complting the network outplts,an error is calculated for all
contnitions to the current practices of neurocontrol. The outputs by comparing the ANN output to a desired output.
next few sections present the neural network and neuro- Reference 15 documents the equations necessary for the
control concepts used in this study. implementation and t d g of the ANN.

2.1 Artifcial Neural Networks 2.2 Supervised learning using backpropagation (BP)of
The ANN structure used in this study is shown in error
Figure 1. It is assumed that every neuron takes connections The most important concept that underlies the ANN
from any or a l l neurons to the left of it. The degree of design is the come@of learning from experience. There are
connectivity depends on the number of connections within several useful learning paradigms for ANN, but the most
the structure. A fully forwardconnected network takes all widely used paradigm is the back-propagation (BP)
possible connections. This ANN configuration provides a algorithm. In BP, the ANN learns by repeated exposure to
network that can have up to "h" (number of hidden neurons) a set of training examples. The leaming takes place through
hidden layers. It is known that as the number of layers a first-order reduction in the output error quantity.
increases the accuracy of the mapping increases. One Backpopgation has been successfully used for mapping of
drawback of increasing the number of layers is that the non-linear functions.
number of parameters to be optimized increases and thus
affects the generalization capabilities of the networks (see 2.3 Neuro-control Techniques
hhelevich et al. [l?). A technique is outlined in reference Two neuroant1-01 techniques used in this study are
15 that optimizes the connectivity pattern as part of the outlined below.
supervised leaming. This technique has the flexibility to off-linesumised controller synthesis using batch learning
maintain the desired accuracy as well as reduce the number Supervised neurocontrol is s& to the indirect
of parameters m the network. Also, the network optimization control used in the adaptive control literature. In traditional
can potentially have up to 'h" neurons in the first layer. It adaptive control, a hear system model is identified and
has been shown before that this is important for using this model the control is adapted. Similarly, in
approximation capabilities of the neural network. As shown supervised neurocontrol, first the system is copied using a
later,tbe opcimauy connected network provides a foundation neural network. This system is called the neuro-model. For
for synthesizing robust neurocontrollers. robust control applications, it is important that the NNM is
general enough to accommodate uncertainties in the system
t modeling. The technique used in this study to arrive at a
U
I
{ Y robust NNM is outlined in Reference 15. (Henceforth,
neural networks whose connections axe pruned to provide
robust mapping are referenced as optimallv connected
networks.) Next, the controller weights are tuned using the
back-propagation of error through the neuro-model. Once
again, it is desired to train a NNC to be general enough to
accommodatemodeling errors. The role of the neuro-model
Figure 1. Artificial Neural Network Structure is to relate the output error to the controller error. It should

1154
be noted that the error definition for the model and the p t control actio-. One approximate, but effective way to
controller are different as shown in Figure 2. The error in simulate this correlation in real-time is to introduce a
the case of the controller is the error between the actual memory filter between the NNM and the error definition (as
system trajectory and a desired trajectory generated by a shown in Figure 3). The bandwidth of the memory filter is
reference model. set less than or equal to the bandwidth of the system. In this
manner, the BP of error "remembers" the past error and
tunes the weights, on-line, proportional to both present and
- - past errors. In this study, a first order filter of the form
I, U(K.1)
~ YAK'
J ~ + ~ = ~ J is~Used.
+ Eh ~ this equation, is the error
S Y m

1 between the desired response and the true response and J is


ii I I the error used to back-propagate through NNM and NNC.
For the memory filter, "a=O" implies a pattern learning and
"a= 1" approximates batch learning.

Figure 2. Neurocontrol using supervised learning and


back-propagation (BP)of error.

For training the NNC, either a batch learning


method or BTT can be used. In this study, a batch learning
as outlined in reference 16 is used. In this method, first the
Figure 3. Neurocontroller training with a memory filter.
system is subjected to an arbitrary disturbance. A forward
pass is performed for each time step without any
backpropagation of error being allowed; however, the total 3.0 NEURO-CONTROL OF SS
error for a time period is collected. During a second
forward pass, the total error from the first pass is Altitude Control of Space Station is challenging due
back-propagated for each time step through the NNM and to the varying mass properties, which change the system
the NNC. Note that the weighted connections of the NNC characteristics continuously. One of the proposed control
are updated during error backpropagation, whereas NNM arclukcams for the rdtihde corprol of Space Station consists
connection weights remain unchanged. of an outer loop for the momentum management and an
inner loop for the attitude control (see Figure 4). The
On-line neuro-controller adaotation using a memory filter: momentum management loop commands the Space Station
On-line adaptation includes learning utunodeled anitude and the inner loop controls the Space Station to
achieve the desired trajectory. A combination of model
system dynamics, persistent system disturbances, and any
other d e s that were not accounted for in the modeling identification and adaptive control can be utilized for the
process. For obvious reasons, batch processing cannot be inner attitude control loop and the reference trajectory is
used in this situation. On the orher hand, BTT requires the generated by the mo"management system. The s d y
error information to flow backwards from time step NT to presented in this paper, without any loss of generality,
time NT-1,to NT-2, and so on. This is inconsistent with any examines only the inner loop pitch attitude control.
ANGLES AND

true real-time adaptation. In this paper, we introduce the RATES

concept of a memory filter that in effect produces a time I I


correlation of the output error for the back-propagation I. U SPACE
process. This approach is outlined below. NNC h'NM

i+r
CONTROL
For on-line adaptive control, the NNM is updated TORQUE
each time step and the NNC weights are tuned using the
updated NNM. In general, the NNM is updated at a
frequency higher than the NNC updates. For the NNC
1 ' INNER LOOP
-
ANCiLLa AND
RATES I
daptation, as statecl earlier, BTT is inadequate for real-time MOMENTUM
applications. A controller action at the present time OUTER
MAh'ACER
influences the future response of the system. The time LOOP U
horizon of this response is directly related to the bandwidth
of the system that is being controlled. BTT essentially Figure 4. Inner and outer loops for Space Station control.
correlates this timedependent error characteristics with the

1155
3.1 Space Station Model
Reference 1 presents the complete non-linear
model of the SS. Assumpt~onsof small rolvyaw attitude
..* .. e,+a,e,=o
generator

ed+a2ed+al
of the form

errors and small products of inertia lead to a simpliiication


of the complete non-linear model. These equations are where s,a,, and a, are chosen to give a characteristic
useful when there is a ueed for large pitch (83maneuvers equation associated with the command generator system as
with small roll (e,) and yaw (83 maneuvers. With the above ( s + b )(s2+2f,0,s+w,3=0 with A=0.0008; <=0.707;and
w,=af5. The values for 1, 5 , and onwere obtained From
simplifiications, the spacecraft's pitch angle dynamics
w o n reduces to: reference 7. The unnmad geoeretor is synmymous with the
outer-loop momentum management presented earlier.

- . i
where €$is the relative pitch angle with respect to the Local-
U
D
Q
.
.
-
-
__-
Descred Output
Opiimally Connected N e i v o r k Output
V d d and Local-Horizontal (LVLH)ecphibrium point, n -2.00 -
is the orbital angular velocity, I,,,, are the principal
of inertia, and 4 is the control torque.

n = 0.0011 d s e ~

I, =5O.28x1O6 slug f?
4= lO.8OxlO6 slug f?
&=58.57~106slugftz 1 w 2.00 3 00 4.3c 5 oc
No. oi Orbits

3 3 off-line Supervised Neuroantroller Synthesis


A "xQ&Olle.r synthesis using a trained neuro-
model is demmdmted in this section. Referring to figure 2, Figure 5 . Comparison of SS and Nemo-model responses
a n m r o - d l is synthesized first using back-propagation of to a random input at the control torque y.
error. The structure of the newal network employed for the
ANN modeling consisted of four inputs (bias,8, and its The structure of the neural network employed for
rate, a n d q at time t), twenty hiddenunits, and two outputs the NNC had two feedback laplts,twelve hidden units, and
(e,amdasrateattimet+dt; dt=100 SWDCIS). ~noptima~y a single output. The two inpd units were the pitch angle
d ANN represedptionusing the technique presented error (e,-eJ d its rate and the smgle output representedthe
m h 15 wasdliztdfor deahfyiug the "M.In this control signal, y Again, only the pitch degree-of-freedom
stxdy, theiaplas a d outputs were linearly scaled so that the was COILsidered for the1- design and verification.
inpdoutpt hyperspace becomes a hyper-cube (equal The NNC was trained using batch learning. The total time
ma@udes fordlinpds and outputs). An appropriate range used for each batch was 15,000 seconds (approximately
of inplts and outputs are chosen for the scaling. For all three orbits).
cases inchded m this study, the inputs and outputs were
scaled to fall between -1 and +l. Reasonable deviations
outside this range are acceptable due to the generalization
capabilities of the ANN. -
, . 10.00 1

The training data for the ANN basically consisted


of rsndam ANN inplts a d the desired response was derived
v 1- Commondea trqectory
Neuro-control trolectory

from the non-linear SS model presented earlier. The error


w a s d d i d as the q " h c error between the ANN outputs
and the corresponding SS model outputs. The computed
error for each time step was back-propagated to update the
weights of the NNM. Figure 5 wmpnx the response of the
SS to that of the trained NNM for a random disturbance
introduced through the control torque, 4.
0
Controller design for the SS involves the training
of a neurocontroller (NNC) to minimize pitch attitude
response re"gfrom any disturbance. The desired output
becomes not the system dynamic response as in model
training, but instead, a desired output magnitude. In this Figure 6. Performance of the neurocontroller for an
study, the desired response was generated by a command initial condition disturbance.

1156
It was originally believed that it would be necessary connected networks, were tested for their robustness to
to train the controller for many random excitations to uzlcertaioees inthemoment of iner&ia of the pitch axis of the
gene& the controller characteristics. h interesting SS. Figures 7 and 8 SUrmDarize the tracking error
dkoveay was ideotified zeganimg the training of a NNC for performances and control energy utilized for control,
control. It WBS onty necessary to train the NNC for a single respectively. It is interesting to note that the optimally
distuhance magnitude and duration. Other disturbance COIlDected network outperforms the fully connected network.
magnitudes and time durations could be controlled without Also, beyond 15 % inertia variation, the controller
having to train the NNC for many random combinations. performances are unacceptable. This is where we need to
Concisely, the NNC had "learned" the dynamics of the adapt the NNM and the NNC to effectively realize an
model and to control the response to any disturbance. Figure adapwe controuer. Figures 7 and 8 also present dah for the
6 shows the response comparison between the neuro- adaptive controller perfonnance for the varying inertia
controller and the c o d e d output for an initial condition characteristics. The adapive control problem is discussed
disturbance. next.

600.00 13.00

j dYVIWWAdoptrve neu-a-control v
--_
d -3

'="-1300 '
:
.
/' - Commended trajectory
Non-adoptive neuro-contro;
Adaptive neuro-control

E -30 03 '
U

0
'

. 0
No. of Oroits
0
i
0

Figure 7. Robustness verification: squared tracking error


for (radian) and 8, ( d s e c ) summed over 15,000
Seconds.

_-- - i
2 -60.00 1
e
- Non odopt ive neuro-con! rol
Adoptive neuro-control

0"c -a0 oo
OOC
3
*
0.50 1.00 1.50 2.30 2.50 3.00 3.53 2.00
No. of Orbiis
IO
2ercent Voriciion in Pitch Inertia

Figure 9. Performance comparison between adaptive and


non-adaptive neuro-control: (a) pitch angle response; and
Figure 8. Robustness verification: squared control effort (b) control torque response.
(4, ft-lb)summed over 15,000seconds.
3.4 Adaptive Neuro-control of SS
For de"g the adapting capability of neuro-
3.3 Robustness of the Neuro-controller controllers, the pitch axis moment of inertia of the SS was
In this phase of the study, the claim that the reduced by 20 A. The adaptation of NNM and NNC were
robustness of the neuro-controller improves if opfimally carried out slsluitaneously (i.e., the weights of both NNM
connected networks are used is verified. Two systems, one and NNC were adapted simultaueously each time step) for
with fully connected networks a d the other with optimally the same initial condition disturbance used in the previous

1157
case. For the BP of error, a first order memory filter is [3] Byun, K. W.,Wie, B., Geller, D., and Sunkel, J. W., 'Robust
H-Infinity Control Design for the Space Station with Structured
introduced between the NNM and the error as shown in
Parameter Uncertainty,' presented at ALAA Guidance, Navigation and
Figure 3. For the first order filter, a time constant of lo00 Control Conference, Portland, OR, August 2@22, 1990.
seconds was used in this study. During the NNC weight 141 Balas, G., hckard, A., and Harduvel, J., 'Applications of
updates, NNM connection weights remain unchanged. Mu-Synthesis Techniques to Momentum Management and AtLitude
Figures 9a and 9b show controlled responses using adaptwe Cor~mlof the Space Station,' presented at AlAA Guidance, Navigation
and Control Conference, January 1991.
a d " - a d a p e -trol. Although the pitch response [SI parks, Akzander G, and Sunkel, J. W., "Adaptive Anitude Control
looks identical, the controller response (uJ for the non- aod Momentum MIM~CUIC~I for Large-Angle Spacecraft Maneuvers, "
adaptive case is unacceptable. For adaptive neuro-control, Jouranl of Guidance, Control, and Dynamics, Vol. ISr, no. 4, July -
the contml response initially starts out oscillatory and 8s August 1992.
[6] Vadali, S. R. and H.S. Oh, "Space Station Attitude Control and
adaptation progresses, the control response improves M o m e m " Management: A nonlincar Look,' Journal of Guidance,
dramatically. Control, and Dynamics, May-June 1992.
4. CONCLUSIONS [7] Bossart, Theodore C. and Singh, Sahjendra N., "Invertibility of
Map, Zero Dynamics and Nonlinear Control of Space Station,' AIAA-
This study demonstrated the applications of neuro- 91-2663-CP.
[8] Parlos, A. G., Atiya, .". F., and Sunkel, I. W., 'Parameter
wotrol to adapve non-Iinear control of Space Station. Two Estimation in Space Systems using Recumnt Neural Networks,' paper
capabilities of neuro-controllers were demonstrated using a no. 91-2716, AIAA Guidance, Navigation, and Control Conference,
non-hear model of the Phase I Space Station Freedom. August 199I .
These capabilities are: (a) synthesis of robust non-linear [9] Chipman, R., Lam, Q., and Sunkel, J. W., 'Mass Property
Identification: A Comparison Between Extended Kalman Filter and
control laws using neural networks; and (b) adaptively
Neurof3ter Approach,' paper no. 91-2664. AIAA Guidance,
"gneurocontrollercharacteristics for varying inertia Navigation, and Control Conference, August 1991.
characteristics using a memory filter. [lo] Kumar, R. R., Seywald, H,Dcshpande, S. M., and Rahman, Z.,
The "-controller techniques presented seem to 'Artificial Nwral Networks in Space Station Optimal Attitude Control, "
be robust based on the successful results obtained using prrsentcd at the World Space Congress, Washington, DC,Aug 28-Sep
5,1992.
straight forward applications of these techniques. To build [ I l l Narendra, K. S., "Adaptive Control Using Neural Networks".
upon the ideas preseded, "-controllers for the three-axis Neural Networks for ControUeditcd by W. Thomas Miller IJI, Richard
w e neuro-controller synthesis is currently in progress. S. Sunon, and Paul J. Werbos. MTT Press, p. 287, Cambridge, MA,
This study examined decentralized neuro-control to 1990.
(121 Werbos, Paul J. "Rack-propagationThrough Time:What It Does
demoostrate the leaming and adaptation capabilities of ANN and How To Do It,.' Proceedings of the EEE, August, 1990.
based on direct ouput fedback. For space applications, the [13] Werbos, P. J. 'A Menu of Designs for Reinfo"t Learning
fixed software/hardware environment provided by ANN Over Ti,' Neural Networks for Control, edited by W. Thomas Miller
synthesis makes it easy to mod@ the controller 111, Richard S. Sumn, and Paul J. Werbos. MIT Ress, p. 67,
characteristics from a remote site, if there is a need. Any Cambridge, MA, 1990.
[14] Barto, A. G., Sumn, R. S., and Anderson, C. W. 'Neuron-like
other controller design could easily be copied using ANN Adaptive elemelll~that can solve difficult control problems", IEEE
a d can take advantage of having a fixed softwarefbardware Trans, SMC., SMC-13, p. 834-846, 1983.
structure for the neuro-controller. [15] KrishnaKumar, K., 'Optimization of the neural net connectivity
Further research in this area needs to examine the pancm using a back-propagation algorithm," Journal of
Neumomputing, 1993.
stabllay and performance robustness of these techniques. [I61 Qin, Si-Zhao, Su, Hong-Te, McAvoy, Thomas J., "Comparison
Specifically, it will be interesting to examine the theoretical of Four Neural Net Learning Methods for Dynamic System
aspects of the optimally connected networks in providing Identification', IEEE Transactions on Neural Networks, Vol. 3, No.
robust control capabilities. Also, the feasibility of using the 1, pp. 122-130, January, 1992.
"oxy fiber concept for dynamic systems with high modal
[I71 V. V. Anshelevich et al., On the ability of neural networks to
perform generalimion by induction, Biological Cybernetics 61, 125-
densities need to be examined. 128, 1989.

ACKNOWLEDGEMENT
This material is based upon work partly supported
by the National Science Foundation under Grant No. ECS-
9113283.

REFERENCES
[ I ] Wie, B., Byun, K. W., Wamn, W., Geller, D., Long, D., and
Sunkel, J. W., 'New Approach to AttitudelMomentumControl of the
Space Station,' Journal of Guidance, Control, and Dynamics, vol 12,
no. 5 , pp. 714-722, 1989.
121 Yeichner, J., Lee, J., and Barmws, O., 'Overview of Space
Station Attitude Control Syacm with Active Momentum
Management,' AAS paper 88-044, February 1988.

1158

You might also like